1
Machine Learning Over Our Growing
Electronic Health Records
Session 165, February 13, 2019
Dr. Kevin Ross, CEO, Precision Driven Health
2
Kevin Ross, Ph.D.
CEO of Precision Driven Health, a research partnership funded
primarily by Orion Health and the New Zealand Government
Conflict of Interest
3
What role can machine learning play in healthcare?
Examples of machine learning
Some data challenges to solve
Building solutions together
Agenda
4
Learning Objective 1: Recognize how machine learning models
can help clinical practice to become personalized care
Learning Objective 2: Recognize the role that data plays in
decision support: the breadth of data we need to understand, and
the scale of records we need to study
Learning Objective 3: Identify how to engage the clinical
community with data scientists to co-develop solutions
Learning Objectives
5
More data, better data science
Rising health costs
Precision health utilises all available
data through machine learning
Healthcare delivery is changing
1
IBM Watson Fortune 2016
Social Data
Exposome
Device Data
Transcriptome
Proteome
Epigenetic
Metabolome
Microbiome
Genome
Imaging
Clinical Data
6Tb of
Data
1
6
Clinicians provide the best possible care, using all
available information.
It is estimated that by the year 2020, it will only take
73 days for the volume of medical knowledge to
double.
How can clinicians keep up?
Our clinicians are amazing data processors
7
Healthcare now
Clinician Patient
Medical Knowledge
Research
Medical
records
8
Healthcare of the future
Clinician
Patient
Medical Knowledge
Research
Medical
records
Machine
learning
9
Machine learning is the use of
algorithms to learn from
different data types, identify
patterns and make predictions.
10
A few examples
Triaging specialist referrals
Searching medical records
Medicines reconciliation
11
Reviewing patient files takes time that could be
spent with patients
Machine learning can accelerate this process,
identifying patients with highest needs
New Zealand: Automating referrals to aid
cardiologists in triaging patients, treating urgent
cases first
Prioritizing patients
12
Clinicians need tools to find clinically meaningful
information
We need to search multiple types of data, for semantic
relationships
Most relevant information is unstructured
Why are health records difficult to search?
13
Semantic search
14
Medicines reconciliation currently a very labour intensive
and error prone task
Medication names and dosage levels need to be
structured
Machine learning can automate this task
What medication is this patient taking?
15
Patient medication timeline
16
Data handling
- Access to private data
- Missing data
- Bias
Models in practice
- Machine learning interpretability
- Communicating
These are crucial but difficult building blocks towards
precision health
We need to solve a few data problems
17
Pharmacy
Community
Safely accessing healthcare data
Meaningful
insights, new
knowledge
Data scientist
Electronic
health records
Consumer
Lab
Payer
Clinical Social
Machine learning
18
HIPAA safe harbour is not anonymization
Simply removing identifiable demographics doesn’t
remove re-identification risk
We are developing a framework from raw data right
through to extracting meaningful insights
De-identification of data
19
Safely accessing healthcare data
De-identification
of records
Electronic
health records
Consumer
Lab
Payer
Clinical Social
Data scientist
Machine learning
Meaningful
insights, new
knowledge
Pharmacy
Community
20
Missing data in
healthcare
Lots of relevant data is missing
Absence may be meaningful
Data not all collected
Our clinicians make decisions with
what data they have, not what they
want
Complete data allows for:
Better representation of groups
in a population
More reliable predictive models
21
How do we handle bias?
Inequity in healthcare are systemic and closely
intertwined with social inequalities.
Research
1
indicates that healthcare professionals
exhibit the same levels of implicit bias as the wider
population.
Bias can be
Source level: how data is generated
Data level: what features are included/excluded
Model level: how the model is trained
1. FitzGerald C, Hurst S. Implicit bias in healthcare professionals: a systematic review. BMC Medical Ethics. 2017;18:19. doi:10.1186/s12910-017-0179-8.
22
Responsible use of data
23
Models must be interpreted to be used
24
Communication of outputs
25
Healthcare of the future
Clinician
Patient
Medical Knowledge
Research
Medical
records
Machine
learning
26
Please complete online session evaluation
Questions
precisiondrivenhealth.com
info@precisiondrivenhealth.com @healthprecision